[abrichr]

Fixes the admission block in the production lifecycle run (packages/flow/benchmark/lifecycle/results/production-2026-10-25-final-blocked/README.md, commit 84ce8bd on measure/production-lifecycle-final).

Before

#1504 skips the candidate census for a target that teach time recorded as a form field or control, but only when the structured tier verified the step. On OpenEMR, step_016 is the "Contact" section button (structural role button), and its pixel crop verifies it. The OCR census had no band to count, and a save follows, so the census turned the verified match into identity_uncounted. The closed-loop scroll before it (step_015) read that as "anchor not in view" and spent its whole budget. The representative case halted in 2 of 2 attempts, so the production program couldn't be admitted. It admitted at 38fc21a.

After

• identity_census.teach_time_role(step) says what teach time recorded:
  • row when the recorder saw a record row (structured_row true);
  • control when it saw none (structured_row false), or when the structural locator names a form-field or control role (textbox, searchbox, combobox, checkbox, radio, spinbutton, switch, slider, button, tab, menuitem);
  • unknown otherwise.
• The census treats control as not a row on every rung (structured, template, OCR, pixel), so a verified identity stands. row and unknown are still counted, so a pixel-resolved link, cell, list item or row, or a target with no recorded role, still halts when it can't be counted. A teach-time row wins over a control role (a row's own button is still counted).
• If the next target is in view but the census refuses it, the closed-loop scroll stops with a safety halt that names the census reason, and sends no input. It no longer scrolls until the budget runs out.

Tests

tests/test_census_teach_role.py:

• A pixel-resolved button, text box, combo box or check box before a save verifies (census_status="not_a_row").
• A pixel-resolved target with no role, or a row, list item, link or grid-cell role, still halts with identity_uncounted.
• A row recorded at teach time with a button role is still counted.
• Two OCR rows that read the recorded band halt with identical_candidates.
• A scroll whose in-view target is refused stops without wheel input, and its error names the census reason.

Locally (1 worker): ruff, mypy (including strict), and 13 test files, 376 passed.

I haven't rerun OpenEMR here; that's the lifecycle run's job.

🤖 Generated with Claude Code

[abrichr]

Real-app check, run by the production-lifecycle session: the OpenEMR representative admission case is VERIFIED with the census on (exact_record read-back, 0 model calls, 108 s engine). It was tested at 76ad93efa: measure/production-lifecycle-final (origin/main 6dc0be4 + #1516 + harness) merged with this branch's 39daa43. Before the fix, the same case halted at step_015 in 2 of 2 attempts. Only the representative case ran, not the 5 fault cases.

[abrichr]

AI review (coordinator)

Verdict: this reopens the wrong-row bug on two paths. I turned off auto-merge. A fix with regression tests is on branch review/1557-census-control-fix (commit f911e85, based on this PR's head).

How teach time decides structured_row

The recorder sets structured_row only when the target has a structured identity and the backend returned a census (recorder.py:695). The browser census finds a row only through closest('tr, [role="row"], li, [role="listitem"]'). Otherwise structured_row stays None. That happens with:

• div, CSS-grid or card rows (no structured identity is recorded, so the census never runs)
• every native backend (Windows, macOS, Linux). They record role="button" but have no identity_candidates_at.
• legacy bundles, and a teach-time census that throws

With None, teach_time_role falls back to the structural role. For any per-row button (button, tab, menuitem, and also checkbox for row-select boxes) that fallback returns "control".

Hole 1: the role skips an OCR census that could count

When the check is not structured, this PR returns not_a_row for a "control" role without running _ocr_census. Before this PR, that census found look-alike rows and halted. I tested a per-row "Open" button with a recorded OCR band and two OCR rows with the same text, at irreversible risk:

role	base (merge-base)	this PR
none	halts, identical_candidates	halts
button / menuitem / tab	halts, identical_candidates	clicks, census_status=not_a_row

Hole 2: a structured row identity with no teach-time census becomes a control

Real headless Chromium. The page has two identical <tr> rows, each with <button id="edit-N">Edit</button>, where the id follows the row's position. I taught "Edit on chart 1001, then Sign" and replayed with the rows swapped, using a backend without a census (_CannotCountBackend, which stands in for native and remote surfaces) for both teach and replay.

• Teach recorded structured_identity='John Smith Follow-up Dr Lee' (that is row text), structured_row=None and role button, so teach_time_role returned control.
• Base: the step halts with identity_uncounted.
• This PR: success=True, status Opened chart 1002, Signed chart 1002, census_status=not_a_row. That is the wrong record.

What still holds

• On Playwright, a <tr> with a button inside gets structured_row=True, so the role is row and the step halts (identical_candidates). This PR doesn't change that.
• The role+name structural locator refuses div rows whose buttons all have the same name ("Open" x3, candidate_count=3). The census isn't involved there.

Not caused by this PR, but open

Div rows whose buttons have positional ids (#open-0) sign the wrong chart on base and on this PR. No identity is armed for that step, and the DOM census can't see div rows, so #1504 never covered them. This needs its own issue: count repeated sibling structures, not only tr/li/role=row.

Fix (on the branch)

--- identity_census.py  teach_time_role
     if recorded_row is False:
         return "control"
+    if getattr(template, "structured", None):
+        # A structured identity was recorded but nothing counted its rows (a
+        # backend without a census, a legacy bundle): it may be a row's text.
+        return "unknown"
--- replayer.py  _apply_candidate_census
         if check.mode == "structured":
             decision = self._structured_census(step, resolution.point)
-            if decision.verdict == "cannot_count" and teach_role == "control":
-                decision = census_mod.CensusDecision("not_a_row")
-        elif teach_role == "control":
-            decision = census_mod.CensusDecision("not_a_row")
         else:
             decision = self._ocr_census(step, resolution, before_png, params, workflow)
+        if decision.verdict == "cannot_count" and teach_role == "control":
+            # A census that can count always decides.
+            decision = census_mod.CensusDecision("not_a_row")

The fix follows one rule: a control role can only replace cannot_count, and a census that can count always decides. The OpenEMR step_016 case still passes. Its pixel-only _ocr_census returns cannot_count (no band), and the role then turns that into not_a_row. All 13 of this PR's tests pass with the fix.

Tests on the branch

New tests:

• test_a_control_role_never_skips_an_ocr_census_that_can_count (button, menuitem, tab, checkbox)
• test_a_recorded_structured_identity_with_no_teach_census_is_not_a_control
• test_realism_browser.py::test_a_row_button_taught_without_a_census_halts_after_a_swap (real Chromium)

Results:

• On this PR's head, all 6 new tests fail.
• With the fix, test_census_teach_role.py, test_realism_browser.py and test_identity_census.py pass, and ruff is clean.

If you merge the branch into this PR, re-enable auto-merge.